Information processing apparatus, information processing system, and information processing method

Abstract

An information processing apparatus that executes machine learning by a plurality of processing layers includes a processor and a memory, a normal region of the memory holds a parameter of the machine learning, an isolation region isolated from the normal region of the memory holds input data to an i-th layer included in the plurality of processing layers and a parameter of the machine learning, the processor determines whether there is a risk in executing processing of the i-th layer for the input data in the normal region based on a content of secret information of the input data in the isolation region, when determining that there is the risk, executes the processing of the i-th layer for the input data in the isolation region, and when determining that there is no risk, executes the processing of the i-th layer for the input data in the normal region.

1 . An information processing apparatus that executes machine learning by a plurality of processing layers, the information processing apparatus, comprising:
 a processor; and   a memory,   wherein   the memory includes a normal region and an isolation region isolated from the normal region,   the normal region holds a parameter of the machine learning,   the isolation region holds input data to an i-th layer included in the plurality of processing layers and a parameter of the machine learning, and   the processor
 executes semantic determination processing of determining whether there is a risk in executing processing of the i-th layer for the input data in the normal region on a basis of a content of secret information of the input data in the isolation region, 
 when determining that there is the risk, executes the processing of the i-th layer for the input data on a basis of the parameter held by the isolation region in the isolation region, and 
 when determining that there is no risk, outputs the input data to the normal region, and executes the processing of the i-th layer for the input data on a basis of the parameter held by the normal region in the normal region. 
   
     
     
         2 . The information processing apparatus according to  claim 1 , wherein the secret information of the input data includes at least one of personal information of the input data and a trade secret of the input data. 
     
     
         3 . The information processing apparatus according to  claim 1 , wherein
 the isolation region holds initial data input to a first layer of the plurality of processing layers, and   the processor,   in the semantic determination processing,
 calculates first similarity between the input data and the initial data, 
 calculates second similarity between the input data and a processing result by a final layer of the plurality of processing layers on a basis of a parameter held by the isolation region, and 
 determines whether there is the risk on a basis of the first similarity and the second similarity. 
   
     
     
         4 . The information processing apparatus according to  claim 3 , wherein
 the information processing apparatus is connected to an input device and a display device, and   the processor
 displays the first similarity and the second similarity on the display device, 
 receives input of a determination result as to whether or not there is the risk via the input device, and 
 determines, as a determination result of the semantic determination processing, the determination result whose input has been received. 
   
     
     
         5 . The information processing apparatus according to  claim 1 , wherein
 the processor   executes quantitative determination processing of determining which of a processing time in a case of executing processing of the i-th layer for the input data in the isolation region or a processing time in a case of executing processing of the i-th layer for the input data in the normal region is shorter, on a basis of an input/output data size of the input data and the parameter held in the isolation region, in the isolation region,   when determining, in the quantitative determination processing, that the processing time is shorter in the case of executing the processing of the i-th layer for the input data in the isolation region, executes the processing of the i-th layer for the input data on a basis of the parameter held by the isolation region in the isolation region without executing the semantic determination processing, and   when determining, in the quantitative determination processing, that the processing time is shorter in the case of executing the processing of the i-th layer for the input data in the normal region, executes the semantic determination processing.   
     
     
         6 . The information processing apparatus according to  claim 5 , wherein
 the isolation region holds
 information indicating first processing time, which is processing time of the machine learning per unit data size in the isolation region, 
 information indicating second processing time, which is processing time of the machine learning per unit data size in the normal region, and 
 information indicating third processing time, which is processing time of the semantic determination processing per unit data size in the isolation region, and 
   on a basis of the input/output data size of the input data, the first processing time, the second processing time, and the third processing time, in the quantitative determination processing, the processor calculates processing time in a case of executing the processing of the i-th layer for the input data in the isolation region and processing time in a case of executing the processing of the i-th layer for the input data in the normal region, and   determines which of the processing time in the case of executing the processing of the i-th layer for the input data in the isolation region or the processing time in the case of executing the processing of the i-th layer for the input data in the normal region is shorter on a basis of the calculated processing time.   
     
     
         7 . The information processing apparatus according to  claim 6 , wherein
 the information processing apparatus is connected to an input device and a display device, and   the processor
 displays the calculated processing time on the display device, 
 receives input of a determination result of the quantitative determination processing via the input device, and 
 determines, as a determination result of the quantitative determination processing, the determination result whose input has been received. 
   
     
     
         8 . An information processing system comprising:
 a plurality of first information processing apparatuses that execute machine learning by a plurality of processing layers; and   a second information processing apparatus,   wherein   the memories included in the plurality of first information processing apparatuses include normal regions and isolation regions isolated from the normal regions,   each of the isolation regions holds input data of the machine learning,   the second information processing apparatus
 holds a parameter of the machine learning, and 
 encrypts and transmits, to the plurality of first information processing apparatuses, the parameter, 
   each of the plurality of first information processing apparatuses,   for each of the plurality of processing layers,
 decrypts and stores, in the normal region and the isolation region, a received encrypted parameter, 
 executes semantic determination processing of determining whether there is a risk in executing processing of the layer for the input data in the normal region on a basis of a content of secret information of the input data in the isolation region, 
 when determining that there is the risk, executes the processing of the layer for the input data on a basis of the parameter held by the isolation region, in the isolation region, 
 when determining that there is no risk, outputs the input data to the normal region, and executes the processing of the layer for the input data on a basis of the parameter held in the normal region, in the normal region, and 
 outputs, to the second information processing apparatus, output data on which the processing of the plurality of processing layers has been executed, and 
   the second information processing apparatus aggregates the output data received from the plurality of first information processing apparatuses.   
     
     
         9 . An information processing method by an information processing server that executes machine learning by a plurality of processing layers, wherein
 the information processing server includes a processor and a memory,   the memory includes a normal region and an isolation region isolated from the normal region,   the normal region holds a parameter of the machine learning, and   the isolation region holds input data to an i-th layer included in the plurality of processing layers and a parameter of the machine learning,   the information processing method comprising:   executing, by the processor, semantic determination processing of determining whether there is a risk in executing processing of the i-th layer for the input data in the normal region on a basis of a content of secret information of the input data in the isolation region;   executing, by the processor, the processing of the i-th layer for the input data on a basis of the parameter held by the isolation region in the isolation region when determining that there is the risk; and   outputting, by the processor, the input data to the normal region, and executes the processing of the i-th layer for the input data on a basis of the parameter held by the normal region in the normal region when determining that there is no risk.